Kevin Hartnett has written a very interesting piece about the mistakes that AI makes when evaluating the content of a computer screen image. This matters for many different scenarios such as automated vehicles. Would a human make a different (and better) decision than the computer if the two are given the same information? Hartnett's piece is based on this new academic study.
Hartnett writes:
"Researchers are still trying to understand exactly why computer vision systems get tripped up so easily, but they have a good guess. It has to do with an ability humans have that AI lacks: the ability to understand when a scene is confusing and thus go back for a second glance."
It appears that people are more humble than the computers. The "arrogant" computers are too sure that they know the true data generating process. This logic is identical to the discussion by Lars Hansen and Tom Sargent on robust decision making . Rational expectations theorists posited that we all agree on the correct data generating process. Robust decision making research retreats from this assumption as economic agents recognize that they "know that they do not know" the true data generating process. Such individuals seek out strategies to protect themselves against worst case scenarios even if they have trouble quantifying these risks.
A humble decision maker recognizes when things "don't look right" and she seeks a 2nd opinion or at least another moment of reflection before making a costly decision.
My new book manuscript pursues this theme of self aware decision makers who build in some flexibility into their menu of choices because they recognize that they themselves are not sure about how the world around them is changing over time.
Given that the computers are trained on backward looking algorithm, they are more likely to make a mistake when confronted with an ambiguous setting. This is a version of the "Lucas Critique", when the rules of the game change --- people reoptimize --- are the AI machines able to do so? This may be the key test of the Alan Turing conjecture.